On 20 October, a San Francisco startup called Radical Numerics published a blog post with a claim that would normally arrive with a journal reference: for the first time, viruses designed by an AI model were administered to a human patient.

The viruses in question are bacteriophages — phages for short — the natural predators of bacteria. They cannot infect human cells, and phage therapy has a century of uneven history and a recent revival as antibiotic resistance spreads. The hard part has never been the concept. It is finding a phage that kills the particular bacterium in front of you, which today means screening collections and, when nothing fits, engineering variants by hand.

Radical Numerics' contribution is Omnii-1, a multimodal genome language model. It is pretrained on genomic sequence and multiomics data, then further trained on phage genomes. The detail that matters for design is context: Omnii-1 can generate DNA for a chosen region while conditioning on the sequence on both sides of it, which is what keeps the rest of a phage viable while you rewrite the part that decides what it attacks.

The target illustrates the problem nicely. T7, a well-studied E. coli phage, has a genome of roughly 40 kilobases — more than seven times the 5.4 kilobases of ΦX174 used in the EvoPhage work published in Science earlier this year. Radical Numerics went after gp17, T7's receptor-binding tail fibre: the protein that decides which bacterial surface the phage can grab.

The numbers in the post are the checkable kind. Of designed variants carrying 11 to 15 amino-acid changes from wild type, 69.9 percent were recovered after rebooting, against 4.8 percent for random controls. The advantage held across the whole range tested, from one change to thirty. Of 6,694 variants with measurable activity, 2,161 beat wild type, including variants up to 28 changes away; random controls that beat wild type were never more than five changes out.

Prioritisation looks similarly practical. On two E. coli strains, 97 and 96 of Omnii-1's top 100 picks cleared the activity threshold. On a third strain, where only 4.5 percent of the evaluated library cleared it, 24 percent of the model's top 100 did. For host prediction, holding out an entire E. coli clade of 37 strains, Omnii-1 still reached 91.4 percent of eligible strains with just two recommended phages. On Pseudomonas aeruginosa, adding E. coli interaction data lifted average precision from 0.570 to 0.608, edging past the specialised GenoPHI workflow at 0.581 — whose performance, notably, fell once the cross-species data was added.

Then the clinical claim. A cocktail containing Omnii-1-designed variants eliminated the patient's antibiotic-resistant isolate in laboratory testing, and was administered under a physician-led compassionate-use pathway. Compassionate use is how unapproved therapies reach patients with no remaining options, under regulatory oversight. It is also, by design, not evidence of efficacy.

Radical Numerics says it will not disclose the patient's clinical course or outcome, on the grounds that even limited clinical detail could make the person identifiable. It supplies the standard caveats itself: the therapy is not approved for broader use, and one patient cannot establish safety or efficacy.

The criticism that followed is fair, and should be stated plainly. There is no peer-reviewed paper behind the clinical claim, no clinical detail for independent review, and no way for an outside group to verify that the outcome was what the announcement implies. The laboratory work is documented in more detail than most startup posts, but it is still a company describing its own results.

What makes the milestone real regardless is the direction of travel. A design-to-clinic loop that once required screening environmental samples now has a generative model in the middle, and that model's outputs are being tested in wet labs and, once, in a person. Whether it becomes a repeatable pathway depends on questions the post does not answer: how fast a cocktail can be produced for a new isolate, how often the bacteria evolve around it, and whether anyone outside the company can reproduce the design pipeline.

One patient is a data point, not a trial. But it is the first time the phrase "AI-designed therapeutic" has had a person on the other end of it.

Sources